[rasz]

How would HRTF be 'performed once at the very end of the signal chain'? Dont you have to transform every individual signal/position before mixing? On the other hand I read somewhere ATMOS is encoded as an array of filters with positions, so decoding is merely a fourier decomposition, would love to learn more about that.

[michaelrmmiller]

There are a few different models at play: surround sound like 5.1, 7.1, ambisonics and the 7.1.4 Atmos static bed; and object-based where mono point source sounds are attached to a location. The former traditional models can be interpreted as individual objects positioned at the speaker locations and folded down to stereo passing through the HRTF that way. It’s a mixed signal so it really is at the end of the chain. For object-based, those are more precisely located but have other downsides (e.g. they break our mixing concepts for things like compression and reverb) and each object would need to be upmixed to binaural stereo through the HRTF.

Higher order ambisonics strikes a pretty good balance in terms of spatial resolution while still being a mixed signal. You can then pair it with objects for specific highlights. Atmos is a 7.1.4 static bed plus dynamic objects, so similar idea. In either case, most of these 3D sound systems support very few dynamic objects. For example, Windows Sonic only supports 15 dynamic objects on Xbox: https://docs.microsoft.com/en-us/windows/win32/coreaudio/spa...